How NOVA-STARS actually scores against the closing line, and against the other public models. Measured, not asserted.
Every level on this site carries a NOVA-STARS projection except the NFL, where we removed it on purpose.
We tested the NFL rating out of sample on the 2025 season — seeding from 2024 only, walking the season in date order, and projecting each game from ratings as they stood before kickoff. It finished 0.06 points better than predicting a tie in every single game, with a 95% confidence interval of −0.80 to +0.89. That is not “a little behind the market.” It is no information at all, on the one level that already has the deepest and most heavily modelled betting market in sports.
There is some signal in it — the projections correlated +0.36 with actual results — but they varied by 9.1 points against an actual spread of 14.2, which is a weak signal stated at full volume. We tested scaling it down; that came back null too. Thirty-two teams playing seventeen games each simply does not give a prior-season margin rating enough to work with.
So NFL cards show the market line, every book’s number, and availability — but no projection of ours. Publishing one would imply information we have measured ourselves not to have. The college and Division II/III models stay, because there the model beats knowing nothing by 5 to 8 points and no sportsbook prices those games at all.